Abstract: Groundwater drought could cause non-negligible harm to groundwater system, social industries and ecosystems which depend on groundwater. It is of great significance to accurately predict groundwater drought. As case studies in the typical regions with less human influence in Hunan Province and Sichuan Province in the Yangtze River Basin, groundwater droughts in typical regions are predicted by using the key influencing factors, such as natural net recharge（P-ET）, landuse and landcover, as predictors of gated recurrent unit（GRU）, which is a deep learning model, and support vector machine（SVM）/multiple linear regression（MLR）, which are machine learning models, respectively. It adopts the groundwater storage anomaly（GWSA） derived from GRACE as the output of the models, and consequently calculates the groundwater drought index based on GRACE. In addition, this paper also explores whether meteorological variables affecting P-ET are suitable as input variables. The main conclusions are as follows:（1） In Hunan, a typical region in the Yangtze River Basin, the MLR has the longest forecast period of GWSA, and the forecast period of SVM is the shortest. During the forecast period, GRU has the best prediction performance and SVM has the worst prediction efficiency. In terms of input variable selection, adding meteorological variables that affect PET improves the prediction performance of SVM and MLR models for GWSA, but reduces the prediction performance of GRU model. In terms of groundwater drought prediction, the case in which the MLR model excludes meteorological variables affecting P-ET shows a better ability to capture groundwater drought dynamics.（2） In Sichuan, a typical region in the Yangtze River Basin, the predicted GWSA of all cases in the test period are close to the reference value, except for some peaks that cannot be caught in the middle and late stages. Adding meteorological variables affecting P-ET can improve the prediction performance of the three models for GWSA. In terms of groundwater drought prediction, using SVM model or MLR model and excluding meteorological variables affecting P-ET can achieve relatively satisfactory performance in groundwater drought prediction.